Make model boundaries aware of is_mult

[capelastegui]

When using multiplicative composition, we sometimes get model coefficients that result in values equal or lower than zero. The plot above shows a model ((linear+ramp)*(season_month*calendar_uk)) where negative coefficients in the monthly seasonality model interact in a weird way with the ramp after Jan 2019.

I think the underlying problem is that some models, like model_season_month, calendar_uk or, in general, any dummy model, shouldn't have negative parameters when using multiplicative composition. Negative parameters are fine with additive composition, though.

This is not currently supported by ForecastModel. ForecastModel.f_bounds() lacks an is_mult parameter. We should add one, and add multiplication-specific boundaries when appropriate.